LECTURE 2

Dimensional Analysis, Scaling, and Similarity

1. Systems of units
The numerical value of any quantity in a mathematical model is measured with
respect to a system of units (for example, meters in a mechanical model, or dollars
in a nancial model). The units used to measure a quantity are arbitrary, and a
change in the system of units (for example, from meters to feet) cannot change the
model.
A crucial property of a quantitative system of units is that the value of a
dimensional quantity may be measured as some multiple of a basic unit. Thus, a
change in the system of units leads to a rescaling of the quantities it measures, and
the ratio of two quantities with the same units does not depend on the particular
choice of the system. The independence of a model from the system of units used to
measure the quantities that appear in it therefore corresponds to a scale-invariance
of the model.
Remark 2.1. Sometimes it is convenient to use a logarithmic scale of units instead
of a linear scale (such as the Richter scale for earthquake magnitudes, or the stellar
magnitude scale for the brightness of stars) but we can convert this to an underlying
linear scale. In other cases, qualitative scales are used (such as the Beaufort wind
force scale), but these scales (leaves rustle or umbrella use becomes dicult)
are not susceptible to a quantitative analysis (unless they are converted in some
way into a measurable linear scale). In any event, we will take connection between
changes in a system of units and rescaling as a basic premise.
A fundamental system of units is a set of independent units from which all
other units in the system can be derived. The notion of independent units can be
made precise in terms of the rank of a suitable matrix [7, 10] but we wont give
the details here.
The choice of fundamental units in a particular class of problems is not unique,
but, given a fundamental system of units, any other derived unit may be constructed
uniquely as a product of powers of the fundamental units.
Example 2.2. In mechanical problems, a fundamental set of units is mass, length,
time, or M, L, T, respectively, for short. With this fundamental system, velocity
V = LT
1
and force F = MLT
2
are derived units. We could instead use, say,
force F, length L, and time T as a fundamental system of units, and then mass
M = FL
1
T
2
is a derived unit.
Example 2.3. In problems involving heat ow, we may introduce temperature
(measured, for example, in Kelvin) as a fundamental unit. The linearity of temper-
ature is somewhat peculiar: although the zeroth law of thermodynamics ensures
that equality of temperature is well dened, it does not say how temperatures can
11
12
be added. Nevertheless, empirical temperature scales are dened, by convention,
to be linear scales between two xed points, while thermodynamics temperature is
an energy, which is additive.
Example 2.4. In problems involving electromagnetism, we may introduce current
as a fundamental unit (measured, for example, in Amp`eres in the SI system) or
charge (measured, for example, in electrostatic units in the cgs system). Unfortu-
nately, the ocially endorsed SI system is often less convenient for theoretical work
than the cgs system, and both systems remain in use.
Not only is the distinction between fundamental and derived units a matter of
choice, or convention, the number of fundamental units is also somewhat arbitrary.
For example, if dimensional constants are present, we may reduce the number of
fundamental units in a given system by setting the dimensional constants equal to
xed dimensionless values.
Example 2.5. In relativistic mechanics, if we use M, L, T as fundamental units,
then the speed of light c is a dimensional constant (c = 3 10
8
ms
1
in SI-units).
Instead, we may set c = 1 and use M, T (for example) as fundamental units. This
means that we measure lengths in terms of the travel-time of light (one nanosecond
being a convenient choice for everyday lengths).
2. Scaling
Let (d
1
, d
2
, . . . , d
r
) denote a fundamental system of units, such as (M, L, T) in
mechanics, and a a quantity that is measurable with respect to this system. Then
the dimension of a, denoted [a], is given by
(2.1) [a] = d
1
1
d
2
2
. . . d
r
r
for suitable exponents (
1
,
2
, . . . ,
r
).
Suppose that (a
1
, a
2
, . . . , a
n
) denotes all of the dimensional quantities appearing
in a particular model, including parameters, dependent variables, and independent
variables. We denote the dimension of a
i
by
(2.2) [a
i
] = d
1,i
1
d
2,i
2
. . . d
r,i
r
.
The invariance of the model under a change in units d
j

j
d
j
implies that it
is invariant under the scaling transformation
a
i

1,i
1

2,i
2
. . .
r,i
r
a
i
i = 1, . . . , n
for any
1
, . . .
r
> 0.
Thus, if
a = f (a
1
, . . . , a
n
)
is any relation between quantities in the model with the dimensions in (2.1) and
(2.2), then f must have the scaling property that

1
1

2
2
. . .
r
r
f (a
1
, . . . , a
n
) = f
_

1,1
1

2,1
2
. . .
r,1
r
a
1
, . . . ,
1,n
1

2,n
2
. . .
r,n
r
a
n
_
.
A particular consequence of this invariance is that any two quantities that are
equal must have the same dimension (otherwise a change in units would violate the
equality). This fact is often useful in nding the dimension of some quantity.
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 13
Example 2.6. According to Newtons second law,
force = rate of change of momentum with respect to time.
Thus, if F denotes the dimension of force and P the dimension of momentum,
then F = P/T. Since P = MV = ML/T, we conclude that F = ML/T
2
(or
mass acceleration).
3. Nondimensionalization
Scale-invariance implies that we can reduce the number of quantities appearing in
a problem by introducing dimensionless quantities.
Suppose that (a
1
, . . . , a
r
) are a set of quantities whose dimensions form a fun-
damental system of units. We denote the remaining quantities in the model by
(b
1
, . . . , b
m
), where r + m = n. Then, for suitable exponents (
1,i
, . . . ,
r,i
) deter-
mined by the dimensions of (a
1
, . . . , a
r
) and b
i
, the quantity

i
=
b
i
a
1,i
1
. . . a
r,i
r
is dimensionless, meaning that it is invariant under the scaling transformations
induced by changes in units.
A dimensionless parameter
i
can typically be interpreted as the ratio of two
quantities of the same dimension appearing in the problem (such as a ratio of
lengths, times, diusivities, and so on). In studying a problem, it is crucial to know
the magnitude of the dimensionless parameters on which it depends, and whether
they are small, large, or roughly of the order one.
Any dimensional equation
a = f(a
1
, . . . , a
r
, b
1
, . . . , b
m
)
is, after rescaling, equivalent to the dimensionless equation
= f(1, . . . , 1,
1
, . . . ,
m
).
Thus, the introduction of dimensionless quantities reduces the number of variables
in the problem by the number of fundamental units. This fact is called the Bucking-
ham Pi-theorem. Moreover, any two systems with the same values of dimensionless
parameters behave in the same way, up to a rescaling.
4. Fluid mechanics
To illustrate the ideas of dimensional analysis, we describe some applications in
uid mechanics.
Consider the ow of a homogeneous uid with speed U and length scale L. We
restrict our attention to incompressible ows, for which U is much smaller that the
speed of sound c
0
in the uid, meaning that the Mach number
M =
U
c
0
is small. The sound speed in air at standard conditions is c
0
= 340 ms
1
. The
incompressibility assumption is typically reasonable when M 0.2.
14
The physical properties of a viscous, incompressible uid depend upon two
dimensional parameters, its mass density
0
and its (dynamic) viscosity . The
dimension of the density is
[
0
] =
M
L
3
.
The dimension of the viscosity, which measures the internal friction of the uid, is
given by
(2.3) [] =
M
LT
.
To derive this result, we explain how the viscosity arises in the constitutive equation
of a Newtonian uid relating the stress and the strain rate.
4.1. The stress tensor
The stress, or force per unit area,

t exerted across a surface by uid on one side of
the surface on uid on the other side is given by

t = Tn
where T is the Cauchy stress tensor and n is a unit vector to the surface. It is
a fundamental result in continuum mechanics, due to Cauchy, that

t is a linear
function of n; thus, T is a second-order tensor [25].
The sign of n is chosen, by convention, so that if n points into uid on one side
A of the surface, and away from uid on the other side B, then Tn is the stress
exerted by A on B. A reversal of the sign of n gives the equal and opposite stress
exerted by B on A.
The stress tensor in a Newtonian uid has the form
(2.4) T = pI + 2D
where p is the uid pressure, is the dynamic viscosity, I is the identity tensor,
and D is the strain-rate tensor
D =
1
2
_
u +u

_
.
Thus, D is the symmetric part of the velocity gradient u.
In components,
T
ij
= p
ij
+
_
u
i
x
j
+
u
j
x
i
_
where
ij
is the Kronecker-,

ij
=
_
1 if i = j,
0 if i ,= j.
Example 2.7. Newtons original denition of viscosity (1687) was for shear ows.
The velocity of a shear ow with strain rate is given by
u = x
2
e
1
where x = (x
1
, x
2
, x
3
) and e
i
is the unit vector in the i
th
direction. The velocity
gradient and strain-rate tensors are
u =
_
_
0 0
0 0 0
0 0 0
_
_
, D =
1
2
_
_
0 0
0 0
0 0 0
_
_
.
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 15
The viscous stress

t
v
= 2Dn exerted by the uid in x
2
> 0 on the uid in x
2
< 0
across the surface x
2
= 0, with unit normal n = e
2
pointing into the region x
2
> 0,
is

t
v
= e
1
. (There is also a normal pressure force

t
p
= pe
1
.) Thus, the frictional
viscous stress exerted by one layer of uid on another is proportional the strain rate
and the viscosity .
4.2. Viscosity
The dynamic viscosity is a constant of proportionality that relates the strain-rate
to the viscous stress.
Stress has the dimension of force/area, so
[T] =
ML
T
2
1
L
2
=
M
LT
2
.
The strain-rate has the dimension of a velocity gradient, or velocity/length, so
[D] =
L
T
1
L
=
1
T
.
Since D has the same dimension as T, we conclude that has the dimension in
(2.3).
The kinematic viscosity of the uid is dened by
=

0
.
It follows from (2.3) that has the dimension of a diusivity,
[] =
L
2
T
.
The kinematic viscosity is a diusivity of momentum; viscous eects lead to the
diusion of momentum in time T over a length scale of the order

T.
The kinematic viscosity of water at standard conditions is approximately 1 mm
2
/s,
meaning that viscous eects diuse uid momentum in one second over a distance
of the order 1 mm. The kinematic viscosity of air at standard conditions is approxi-
mately 15 mm
2
/s; it is larger than that of water because of the lower density of air.
These values are small on every-day scales. For example, the timescale for viscous
diusion across room of width 10 m is of the order of 6 10
6
s, or about 77 days.
4.3. The Reynolds number
The dimensional parameters that characterize a uid ow are a typical velocity U
and length L, the kinematic viscosity , and the uid density
0
. Their dimensions
are
[U] =
L
T
, [L] = L, [] =
L
2
T
, [
0
] =
M
L
3
.
We can form a single independent dimensionless parameter from these dimensional
parameters, the Reynolds number
(2.5) R =
UL

.
As long as the assumptions of the original incompressible model apply, the behavior
of a ow with similar boundary and initial conditions depends only on its Reynolds
number.
16
The inertial term in the Navier-Stokes equation has the order of magnitude

0
u u = O
_

0
U
2
L
_
,
while the viscous term has the order of magnitude
u = O
_
U
L
2
_
.
The Reynolds number may therefore be interpreted as a ratio of the magnitudes of
the inertial and viscous terms.
The Reynolds number spans a large range of values in naturally occurring ows,
from 10
20
in the very slow ows of the earths mantle, to 10
5
for the motion of
bacteria in a uid, to 10
6
for air ow past a car traveling at 60 mph, to 10
10
in
some large-scale geophysical ows.
Example 2.8. Consider a sphere of radius L moving through an incompressible
uid with constant speed U. A primary quantity of interest is the total drag force
D exerted by the uid on the sphere. The drag is a function of the parameters on
which the problem depends, meaning that
D = f(U, L,
0
, ).
The drag D has the dimension of force (ML/T
2
), so dimensional analysis implies
that
D =
0
U
2
L
2
F
_
UL

_
.
Thus, the dimensionless drag
(2.6)
D

0
U
2
L
2
= F(R)
is a function of the Reynolds number (2.5), and dimensional analysis reduces the
problem of nding a function f of four variables to nding a function F of one
variable.
The function F(R) has a complicated dependence on R which is dicult to
determine theoretically, especially for large values of the Reynolds number. Never-
theless, experimental measurements of the drag for a wide variety of values of U,
L,
0
and agree well with (2.6)
4.4. The Navier-Stokes equations
The ow of an incompressible homogeneous uid with density
0
and viscosity is
described by the incompressible Navier-Stokes equations,

0
(u
t
+u u) +p = u,
u = 0.
(2.7)
Here, u(x, t) is the velocity of the uid, and p (x, t) is the pressure. The rst
equation is conservation of momentum, and the second equation is conservation of
volume.
Remark 2.9. It remains an open question whether or not the three-dimensional
Navier-Stokes equations, with arbitrary smooth initial data and appropriate bound-
ary conditions, have a unique, smooth solution that is dened for all positive times.
This is one of the Clay Institute Millenium Prize Problems.
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 17
Let U, L be a typical velocity scale and length scale of a uid ow, and dene
dimensionless variables by
u

=
u
U
, p

=
p
U
2
, x

=
x
L
, t

=
Ut
L
.
Using these expressions in (2.7), and dropping the stars on the dimensionless vari-
ables, we get
u
t
+u u +p =
1
R
u,
u = 0,
(2.8)
where R is the Reynolds number dened in (2.5).
4.5. Euler equations
The nondimensionalized equation (2.8) suggests that for ows with high Reynolds
number, we may neglect the viscous term on the right hand side of the momentum
equation, and approximate the Navier-Stokes equation by the incompressible Euler
equations
u
t
+u u +p = 0,
u = 0.
The Euler equations are dicult to analyze because, like the Navier-Stokes
equations, they are nonlinear. Moreover, the approximation of the Navier-Stokes
equation by the Euler equations is problematic. High-Reynolds number ows de-
velop complicated small-scale structures (for instance, boundary layers and turbu-
lence) and, as a result, it is not always possible to neglect the second-order spatial
derivatives u in the viscous term in comparison with the rst-order spatial deriva-
tives u u in the inertial term, even though the viscous term is multiplied by a
small coecient.
4.6. Stokes equations
At low Reynolds numbers a dierent nondimensionalization of the pressure, based
on the viscosity rather than the inertia, is appropriate. Using
u

=
u
U
, p

=
p
U
2
, x

=
x
L
, t

=
Ut
L
,
in (2.7), and dropping the stars on the dimensionless variables, we get
R(u
t
+u u) +p = u,
u = 0.
Setting R = 0 in these equations, we get the Stokes equations,
(2.9) p = u, u = 0.
These equations provide a good approximation for low Reynolds number ows (al-
though nonuniformities arise in using them on unbounded domains). They are
much simpler to analyze than the full Navier-Stokes equations because they are
linear.
18
5. Stokes formula for the drag on a sphere
As an example of the solution of the Stokes equations for low Reynolds number
ows, we will derive Stokes formula (1851) for the drag on a sphere moving at
constant velocity through a highly viscous uid.
It is convenient to retain dimensional variables, so we consider Stokes equations
(2.9) in dimensional form
(2.10) u = p, u = 0.
We note for future use that we can eliminate the pressure from (2.10) by taking the
curl of the momentum equation, which gives
(2.11) curl u = 0.
Before considering axisymmetric Stokes ow past a sphere, it is useful to look
at the two-dimensional equations. Using Cartesian coordinates with x = (x, y) and
u = (u, v), we may write (2.10) as 3 3 system for (u, v, p):
u = p
x
, v = p
y
, u
x
+v
y
= 0.
Here, =
2
x
+
2
y
is the two-dimensional Laplacian. In a simply connected region,
the incompressibility condition implies that we may introduce a streamfunction
(x, y) such that u =
y
and v =
x
. The momentum equation then becomes

y
= p
x
,
x
= p
y
.
The elimination of p by cross-dierentiation implies that satises the biharmonic
equation

2
= 0.
Thus, the two-dimensional Stokes equations reduce to the biharmonic equation.
Similar considerations apply to axisymmetric ows, although the details are
more complicated. We will therefore give a direct derivation of the solution for ow
past a sphere, following Landau and Lifshitz [34].
We denote the radius of the sphere by a, and adopt a reference frame moving
with the sphere. In this reference frame, the sphere is at rest and the uid velocity
far away from the sphere approaches a constant velocity

U. The pressure also
approaches a constant, which we may take to be zero without loss of generality.
The appropriate boundary condition for the ow of a viscous uid past a solid,
impermeable boundary is the no-slip condition that the velocity of the uid is
equal to the velocity of the body. Roughly speaking, this means that a viscous
uid sticks to a solid boundary.
Let x denote the position vector from the center of the sphere and r = [x[ the
distance. We want to solve the Stokes equations (2.10) for u(x), p(x) in the exterior
of the sphere a < r < , subject to the no-slip condition on the sphere,
(2.12) u(x) = 0 at r = a,
and the uniform-ow condition at innity,
(2.13) u(x)

U, p(x) 0 as r .
First, we will solve for the velocity. Since u is divergence free and the exterior
of a sphere is simply connected, we can write it as
(2.14) u(x) =

U + curl

A(x)
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 19
where

A is a vector-streamfunction for the deviation of the ow from the uniform
ow. We can always choose

A to be divergence free, and we require that the
derivatives of

A approach zero at innity so that u approaches

U.
We will show that we can obtain the solution by choosing

A to be of the form
(2.15)

A(x) = f(r)

U
for a suitable scalar valued function f(r) which we will determine. This form for

A is dictated by linearity and symmetry considerations: since the Stokes equations

are linear, the solution must be linear in the velocity

U; and the solution must be
invariant under rotations about the axis parallel to

U through the center of the
sphere, and under rotations of

U.
Using the vector identity
curl
_
f

F
_
= f

F +f curl

F,
and the fact that

U is constant, we may also write (2.15) as
(2.16)

A = curl
_
f

U
_
,
which shows, in particular, that

A = 0.
By use of the vector identity
(2.17) curl curl

F =
_

F
_

F,
and (2.15), we nd that
curl u = curl curl

A =

A =
_
f

U
_
.
Using this result in (2.11), we nd that

2
_
f

U
_
= 0.
Since

U is constant, it follows that

2
f
_

U = 0.
Since f depends only on r, this equation implies that
_

2
f
_
= 0, so
2
f is
constant. Since the derivatives of

A decay at innity,
2
f must also decay, so the
constant is zero, and therefore f satises the biharmonic equation
(2.18)
2
f = 0.
Writing g = f, which is a function of r = [x[, and using the expression for
the three-dimensional Laplacian in spherical-polar coordinates, we get
1
r
2
d
dr
_
r
2
dg
dr
_
= 0.
Integrating this equation, we get g(r) = 2b/r +c where b, c are constant of integra-
tion. Since f 0 as r , we must have c = 0, so
f =
1
r
2
d
dr
_
r
2
df
dr
_
=
2b
r
.
20
Integrating this equation and neglecting an additive constant, which involves no
loss of generality because

A depends only on f, we get
(2.19) f(r) = br +
c
r
where c is another constant of integration.
Using this expression for f in (2.15), then using the result in (2.14), we nd
that
(2.20) u(x) =

U
b
r
_

U +
1
r
2
_

U x
_
x
_
+
c
r
3
_
3
r
2
_

U x
_
x

U
_
.
This velocity eld satises the boundary condition at innity (2.13). Imposing the
boundary condition (2.12) on the sphere, we get
_
1
b
a

c
a
3
_

U +
1
a
3
_
3c
a
2
b
_
_

U x
_
x = 0 when [x[ = a.
This condition is satised only if the coecient of each term vanishes, which gives
b =
3a
4
, c =
a
3
4
.
Thus, from (2.19), the solution for f is
(2.21) f(r) =
3ar
4
_
1 +
a
2
3r
2
_
,
and, from (2.20), the solution for the velocity eld is
(2.22) u(x) =
_
1
3a
4r

a
3
4r
3
_

U +
1
r
2
_
3a
3
4r
3

3a
4r
_
_

U x
_
x.
Remark 2.10. A noteworthy feature of this solution is its slow decay to the uniform
ow. The dierence

U u(x)
3a
4r
_

U +
1
r
2
_

U x
_
x
_
is of the order 1/r as r . This makes the analysis of nondilute suspensions of
particles dicult, even with linearity of the Stokes equations, because the hydro-
dynamic interactions between particles have a long range.
To get the pressure, we rst compute u. Using (2.16) in (2.14), and applying
the vector identity (2.17), we get
u =

U + curl curl
_
f

U
_
=

U +
_

_
f

U
__
(f)

U.
Taking the Laplacian of this equation, then using the identity
_
f

U
_
=

U f
and the fact that f satises the biharmonic equation (2.18), we get
u =
_

U f
_
.
Use of this expression in the momentum equation in (2.10) gives

U f
_
p
_
= 0.
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 21
It follows that the expression inside the gradient is constant, and from (2.13) the
constant is zero. Therefore,
p =
_

U f
_
.
Using (2.21) in this equation, we nd the explicit expression
(2.23) p =
_
3a
2r
3
_

U x.
Thus, (2.22) and (2.23) is the solution of (2.10) subject to the boundary conditions
(2.13)(2.12).
A primary quantity of interest is the drag force F exerted by the uid on the
sphere. This force is given by integrating the stress over the surface of the
sphere:

F =
_

TndS.
Here, n is the unit outward normal to the sphere, and T is the Cauchy stress tensor,
given from (2.4) by
T = pI +
_
u +u

_
.
A direct calculation, whose details we omit, shows that the force is in the direction
of

U with magnitude
(2.24) F = 6aU,
where U is the magnitude of

U.
This expression for the drag on a spherical particle is found to be in excellent
agreement with experimental results if R < 0.5, where
R =
2aU

is the Reynolds numbers based on the particle diameter, and = /

0
is the
kinematic viscosity, as before.
For example, consider a particle of radius a and density
p
falling under gravity
in a uid of density
0
. At the terminal velocity U, the viscous drag must balance
the gravitational buoyancy force, so
6aU =
4
3
a
3
(
p

0
) g
where g is the acceleration due to gravity. This gives
U =
2a
2
g
9
_

0
1
_
The corresponding Reynolds number is
R =
4a
3
g
9
2
_

0
1
_
.
For a water droplet falling through air [9], we have
p
/
0
780 and
15 mms
1
. This gives a Reynolds number of approximately 1.5 10
4
a
3
where
a is measured in mm. Thus, Stokes formula is applicable when a 0.04 mm,
corresponding to droplets in a ne mist.
22
6. Kolmogorovs 1941 theory of turbulence
Finally, if we are to list the reasons for studying homogeneous
turbulence, we should add that it is a profoundly interesting
physical phenomenon which still dees satisfactory mathematical
analysis; this is, of course, the most compelling reason.
1
High-Reynolds number ows typically exhibit an extremely complicated behav-
ior called turbulence. In fact, Reynolds rst introduced the Reynolds number in
connection with his studies on transition to turbulence in pipe ows in 1895. The
analysis and understanding of turbulence remains a fundamental challenge. There
is, however, no precise denition of uid turbulence, and there are many dierent
kinds of turbulent ows, so this challenge is likely to be one with many dierent
parts.
In 1941, Kolmogorov proposed a simple dimensional argument that is one of
the basic results about turbulence. To explain his argument, we begin by describing
an idealized type of turbulence called homogeneous, isotropic turbulence.
6.1. Homogeneous, isotropic turbulence
Following Batchelor [8], let us imagine an innite extent of uid in turbulent motion.
This means, rst, that the uid velocity depends on a large range of length scales;
we denote the smallest length scale (the dissipation length scale) by
d
and the
largest length scale (the integral length scale) by L. And, second, that the uid
motion is apparently random and not reproducible in detail from one experiment
to the next.
We therefore adopt a probabilistic description, and suppose that a turbulent
ow is described by a probability measure on solutions of the Navier-Stokes equa-
tions such that expected values of the uid variables with respect to the measure
agree with appropriate averages of the turbulent ow.
This probabilistic description is sometimes interpreted as follows: we have an
ensemble of many dierent uid ows obtained, for example, by repeating
the same experiment many dierent times and each member of the ensemble
corresponds to a ow chosen at random with respect to the probability measure.
A turbulent ow is said to be homogeneous if its expected values are invariant
under spatial translations meaning that, on average, it behaves the same way at
each point in space and isotropic if its expected values are also independent of
spatial rotations. Similarly, the ow is stationary if its expected values are invariant
under translations in time. Of course, any particular realization of the ow varies
in space and time.
Homogeneous, isotropic, stationary turbulence is rather unphysical. Turbu-
lence is typically generated at boundaries, and the properties of the ow vary
with distance from the boundary or other large-scale features of the ow geom-
etry. Moreover, turbulence dissipates energy at a rate which appears to be nonzero
even in the limit of innite Reynolds number. Thus, some sort of forcing (usually
at the integral length scale) that adds energy to the uid is required to maintain
stationary turbulence. Nevertheless, appropriate experimental congurations (for
example, high-Reynolds number ow downstream of a metal grid) and numerical
congurations (for example, direct numerical simulations on a box with periodic
1
G. K. Batchelor, The Theory of Homogeneous Turbulence.
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 23
boundary conditions and a suitable applied force) provide a good approximation to
homogeneous, isotropic turbulence.
6.2. Correlation functions and the energy spectrum
We denote expected values by angular brackets . In a homogeneous ow the
two-point correlation
(2.25) Q = u(x, t) u(x +r, t)
is a function of the spatial displacement r, and independent of x. In a stationary
ow it is independent of t. Furthermore, in an isotropic ow, Q is a function only
of the magnitude r = [r[ of the displacement vector.
Note, however, that even in isotropic ow the general correlation tensor
Q(r) = u(x, t) u(x +r, t) ,
with components Q
ij
= u
i
u
j
, depends on the vector r, not just its magnitude,
because a rotation of r also induces a rotation of u.
For isotropic turbulence, one can show [8] that the two-point correlation (2.25)
has the Fourier representation
Q(r) = 2
_

0
sin kr
kr
E(k) dk
where E(k) is a nonnegative function of the wavenumber magnitude k.
In particular, it follows that
(2.26)
1
2
u(x, t) u(x, t) =
_

0
E(k) dk.
Thus, E(k) may be interpreted as the mean kinetic energy density of the turbulent
ow as a function of the wavenumber 0 k < .
6.3. The ve-thirds law
In fully developed turbulence, there is a wide range of length scales
d
L
that are much greater than the dissipation length scale and much less than the
integral length scale. This range is called the inertial range. The corresponding
wavenumbers are k = 2/, with dimension
[k] =
1
L
.
It appears reasonable to assume that the components of a turbulent ow which
vary over length scales in the inertial range do not depend on the viscosity of the
uid or on the integral length scale and velocity.
Kolmogorov proposed that, in the inertial range, the ow statistics depend only
on the mean rate per unit mass at which the turbulent ow dissipates energy. It
would not make any dierence if we used instead the mean rate of energy dissipation
per unit volume, since we would have to nondimensionalize this by the uid density,
to get the mean energy dissipation rate per unit mass. The dimension of this rate
is is
[] =
ML
2
T
2

1
T

1
M
=
L
2
T
3
.
From (2.26), the spectral energy density has dimension
[E(k)] =
L
3
T
2
.
24
If the only quantities on which E(k) depends are the energy dissipation rate and
the wavenumber k itself, then, balancing dimensions, we must have
(2.27) E(k) = C
2/3
k
5/3
,
where C is a dimensionless constant, called the Kolmogorov constant.
Thus, Kolmogorovs 1941 (K41) theory predicts that the energy spectrum of a
turbulent ow in the inertial range has a power-law decay as a function of wavenum-
ber with exponent 5/3; this is the ve-thirds law.
The spectral result (2.27) was, in fact, rst stated by Oboukhov (1941). Kol-
mogorov gave a closely related result for spatial correlations:
_
[u(x +r, t) u(x, t)[
2
_
= C
2/3
r
2/3
.
This equation suggests that the velocity of a turbulent ow has a rough spatial
dependence in the inertial range, similar to that of a non-dierentiable Holder-
continuous function with exponent 1/3.
Onsager rediscovered this result in 1945, and in 1949 suggested that turbulent
dissipation might be described by solutions of the Euler equation that are not suf-
ciently smooth to conserve energy [20]. The possible relationship of non-smooth,
weak solutions of the incompressible Euler equations (which are highly non-unique
and can even increase in kinetic energy without some kind of additional admis-
sibility conditions) to turbulent solutions of the Navier-Stokes equations remains
unclear.
6.4. The Kolmogorov length scale
The only length scale that can be constructed from the dissipation rate and the
kinematic viscosity , called the Kolmogorov length scale, is
=
_

_
1/4
.
The K41 theory implies that the dissipation length scale is of the order .
If the energy dissipation rate is the same at all length scales, then, neglecting
order one factors, we have
=
U
3
L
where L, U are the integral length and velocity scales. Denoting by R
L
the Reynolds
number based on these scales,
R
L
=
UL

,
it follows that
L

= R
3/4
L
.
Thus, according to this dimensional analysis, the ratio of the largest (integral)
length scale and the smallest (dissipation) length scale grows like R
3/4
L
as R
L
.
In order to resolve the nest length scales of a three-dimensional ow with
integral-scale Reynolds number R
L
, we therefore need on the order of
N
L
= R
9/4
L
independent degrees of freedom (for example, N
L
Fourier coecients of the velocity
components). The rapid growth of N
L
with R
L
limits the Reynolds numbers that
can be attained in direct numerical simulations of turbulent ows.
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 25
6.5. Validity of the ve-thirds law
Experimental observations, such as those made by Grant, Stewart and Moilliet
(1962) in a tidal channel between islands o Vancouver, agree well with the ve-
thirds law for the energy spectrum, and give C 1.5 in (2.27). The results of DNS
on periodic boxes, using up to 4096
3
grid points, are also in reasonable agreement
with this prediction.
Although the energy spectrum predicted by the K41 theory is close to what is
observed, there is evidence that it is not exactly correct. This would imply that
there is something wrong with its original assumptions.
Kolmogorov and Oboukhov proposed a renement of Kolmogorovs original
theory in 1962. It is, in particular, not clear that the energy dissipation rate
should be assumed constant, since the energy dissipation in a turbulent ow itself
varies over multiple length scales in a complicated fashion. This phenomenon,
called intermittency, can lead to corrections in the ve-thirds law [23]. All such
turbulence theories, however, depend on some kind of initial assumptions whose
validity can only be checked by comparing their predictions with experimental or
numerical observations.
6.6. The benets and drawbacks of dimensional arguments
As the above examples from uid mechanics illustrate, dimensional arguments can
lead to surprisingly powerful results, even without a detailed analysis of the under-
lying equations. All that is required is a knowledge of the quantities on which the
problem being studied depends together with their dimensions. This does mean,
however, one has to know the basic laws that govern the problem, and the dimen-
sional constants they involve. Thus, contrary to the way it sometimes appears,
dimensional analysis does not give something for nothing; it can only give what is
put in from the start.
This fact cuts both ways. Many of the successes of dimensional analysis, such
as Kolmogorovs theory of turbulence, are the result of an insight into which dimen-
sional parameters play an crucial role in a problem and which parameters can be
ignored. Such insights typical depend upon a great deal of intuition and experience,
and they may be dicult to justify or prove.
2
Conversely, it may happen that some dimensional parameters that appear to
be so small they can be neglected have a signicant eect, in which case scaling
laws derived on the basis of dimensional arguments that ignore them are likely to
be incorrect.
7. Self-similarity
If a problem depends on more fundamental units than the number of dimensional
parameters, then we must use the independent or dependent variables themselves
to nondimensionalize the problem. For example, we did this when we used the
wavenumber k to nondimensionalize the K41 energy spectrum E(k) in (2.27). In
2
As Bridgeman ([10], p. 5) puts it in his elegant 1922 book on dimensional analysis (well worth
reading today): The untutored savage in the bushes would probably not be able to apply the
methods of dimensional analysis to this problem and obtain results that would satisfy us. Hope-
fully, whatever knowledge may have been lost since then in the area of dimensional analysis has
been oset by some gains in cultural sensitivity.
26
that case, we obtain self-similar solutions that are invariant under the scaling trans-
formations induced by a change in the system of units. For example, in a time-
dependent problem the spatial prole of a solution at one instant of time might be
a rescaling of the spatial prole at any other time.
These self-similar solutions are often among the few solutions of nonlinear equa-
tions that can be obtained analytically, and they can provide valuable insight into
the behavior of general solutions. For example, the long-time asymptotics of solu-
tions, or the behavior of solutions at singularities, may be given by suitable self-
similar solutions.
As a rst example, we use dimensional arguments to nd the Greens function
of the heat equation.
7.1. The heat equation
Consider the following IVP for the Greens function of the heat equation in R
d
:
u
t
= u,
u(x, 0) = E(x).
Here is the delta-function, representing a unit point source at the origin. Formally,
we have
_
R
d
(x) dx = 1, (x) = 0 for x ,= 0.
The dimensioned parameters in this problem are the diusivity and the energy
E of the point source. The only length and times scales are those that come from
the independent variables (x, t), so the solution is self-similar.
We have [u] = , where denotes a unit of temperature. Furthermore, since
_
R
d
u(x, 0) dx = E,
we have [E] = L
d
. The rotational invariance of the Laplacian, and the uniqueness
of the solution, implies that the solution must be spherically symmetric. Dimen-
sional analysis then gives
u(x, t) =
E
(t)
d/2
f
_
[x[

t
_
.
Using this expression for u(x, t) in the PDE, we get an ODE for f(),
f

+
_

2
+
d 1

_
f

+
d
2
f = 0.
We can rewrite this equation as a rst-order ODE for f

+

2
f,
_
f

+

2
f
_

+
d 1

_
f

+

2
f
_
= 0.
Solving this equation, we get
f

+

2
f =
b

d1
,
where b is a constant of integration. Solving for f, we get
f() = ae

2
/4
+be

2
/4
_
e

d1
d,
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 27
where a s another constant of integration. In order for f to be integrable, we must
set b = 0. Then
u(x, t) =
aE
(t)
d/2
exp
_

[x[
2
4t
_
.
Imposing the requirement that
_
R
d
u(x, t) dx = E,
and using the standard integral
_
R
d
exp
_

[x[
2
2c
_
dx = (2c)
d/2
,
we nd that a = (4)
d/2
, and
u(x, t) =
E
(4t)
d/2
exp
_

[x[
2
4t
_
.
8. The porous medium equation
In this section, we will further illustrate the use of self-similar solutions by describing
a problem for point-source solutions of the porous medium equation, taken from
Barenblatt [7]. This solution is a one-dimensional version of the radially symmetric
self-similar solution of the porous medium equation
u
t
= (uu)
found by Zeldovich and Kompaneets (1950) and Barenblatt (1952).
We consider the ow under gravity of an incompressible uid in a porous
medium, such as groundwater in a rock formation. We suppose that the porous
medium sits above a horizontal impermeable stratum, and, to simplify the dis-
cussion, that the ow is two-dimensional. It is straightforward to treat three-
dimensional ows in a similar way.
Let x and z denote horizontal and vertical spatial coordinates, respectively,
where the impermeable stratum is located at z = 0. Suppose that the porous
medium is saturated with uid for 0 z h(x, t) and dry for z > h(x, t). If the
wetting front z = h(x, t) has small slope, we may use a quasi-one dimensional ap-
proximation in which we neglect the z-velocity components of the uid and average
x-velocity components with respect to z.
The volume of uid (per unit length in the transverse y-direction) in a x b
is given by
_
b
a
nh(x, t) dx
where n is the porosity of the medium. That is, n is the ratio of the open volume
in the medium that can be occupied by uid to the total volume. Typical values of
n are 0.30.7 for clay, and 0.01, or less, for dense crystalline rocks. We will assume
that n is constant, in which case it will cancel from the equations.
Let u(x, t) denote the depth-averaged x-component of the uid velocity. Con-
servation of volume for an incompressible uid implies that for any x-interval [a, b]
d
dt
_
b
a
nh(x, t) dx = [nhu]
b
a
.
28
In dierential form, we get
(2.28) h
t
= (hu)
x
For slow ows, we can assume that the pressure p in the uid is equal to the
hydrostatic pressure
p =
0
g (h z) .
It follows that the total pressure head, dened by
p

0
g
+z,
is independent of z and equal to h(x, t).
According to Darcys law, the volume-ux (or velocity) of a uid in a porous
medium is proportional to the gradient of the pressure head, meaning that
(2.29) u = kh
x
,
where k is the permeability, or hydraulic conductivity, of the porous medium.
Remark 2.11. Henri Darcy was a French water works engineer. He published his
law in 1856 after conducting experiments on the ow of water through columns of
sand, which he carried out in the course of investigating fountains in Dijon.
The permeability k in (2.29) has the dimension of L
2
/(HT), where H is the
dimension of the head h. Since we measure h in terms of vertical height, k has
the dimension of velocity. Typical values of k for consolidated rocks range from
10
9
m/day for unfractured metamorphic rocks, to 10
3
m/day for karstic limestone.
Using (2.29) in (2.28), we nd that h(x, t) satises the porous medium equation
(2.30) h
t
= k (hh
x
)
x
.
We may alternatively write (2.30) as
h
t
=
1
2
k
_
h
2
_
xx
.
This equation was rst considered by Boussinesq (1904).
The porous medium equation is an example of a degenerate diusion equation.
It has a nonlinear diusivity equal to kh which vanishes when h = 0. As we will
see, this has the interesting consequence that (2.30) has solutions (corresponding to
wetting fronts) that propagate into a region with h = 0 at nite speed behavior
one would expect of a wave equation, but not at rst sight of a diusion equation.
8.1. A point source solution
Consider a solution h(x, t) of the porous medium equation (2.30) that approaches
an initial point source:
h(x, t) I(x), t 0
+
,
where (x) denotes the Dirac delta function. Explicitly, this means that we require
h(x, t) 0 as t 0
+
if x ,= 0,
lim
t0
+
_

h(x, t) dx = I.
(2.31)
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 29
The delta-function is a distribution, rather than a function. We will not discuss
distribution theory here (see [44] for an introduction and [27] for a detailed ac-
count). Instead, we will dene the delta-function formally as a function with
the properties that
(x) = 0 for x ,= 0,
_

f(x)(x) dx = f(0)
for any continuous function f.
The solution of the porous medium with the initial data (2.31) describes the
development of of wetting front due to an instantaneous ooding at the origin by
a volume of water I. It provides the long time asymptotic behavior of solutions of
the porous medium equation with a concentrated non-point source initial condition
h(x, t) = h
0
(x) where h
0
is a compactly supported function with integral I.
The dimensional parameters at our disposal in solving this problem are k and
I. A fundamental system of units is L, T, H where H is a unit for the pressure
head. Since we measure the pressure head h in units of length, it is reasonable to
ask why we should use dierent units for h and x. The explanation is that the units
of vertical length used to measure the head play a dierent role in the model than
the units used to measure horizontal lengths, and we should be able to rescale x
and z independently.
Equating the dimension of dierent terms in (2.30), we nd that
[k] =
L
2
HT
, [I] = LH.
Since we assume that the initial data is a point source, which does not dene a
length scale, there are no other parameters in the problem.
Two parameters k, I are not sucient to nondimensionalize a problem with
three fundamental units. Thus, we must also use one of the variables to do so.
Using t, we get
_
(kIt)
1/3
_
= L, [t] = T,
_
I
2/3
(kt)
1/3
_
= H
Dimensional analysis then implies that
h(x, t) =
I
2/3
(kt)
1/3
F
_
x
(kIt)
1/3
_
where F() is a dimensionless function.
Using this similarity form in (2.30), we nd that F() satises the ODE
(2.32)
1
3
(F

+F) = (FF

.
Furthermore, (2.31) implies that
F() 0 as [[ , (2.33)
_

F() d = 1. (2.34)
30
Integrating (2.32), we get
(2.35)
1
3
F +C = FF

where C is a constant of integration.

The condition (2.33) implies that C = 0. It then follows from (2.35) that either
F = 0, or
F

=
1
3
,
which implies that
F() =
1
6
_
a
2

2
_
where a is a constant of integration.
In order to get a solution that is continuous and approaches zero as [[ ,
we choose
F() =
_ _
a
2

2
_
/6 if [[ < a,
0 if [[ a.
The condition (2.34) then implies that
a =
_
9
2
_
1/3
.
Thus, the solution of (2.30)(2.31) is given by
h(x, t) =
I
2/3
6 (kt)
1/3
_
_
9
2
_
2/3

x
2
(kIt)
2/3
_
if [x[ < (9kIt/2)
1/3
with h(x, t) = 0 otherwise.
This solution represents a saturated region of nite width which spreads out
at nite speed. The solution is not a classical solution of (2.30) since its derivative
h
x
has a jump discontinuity at x = (9kIt/2)
1/3
. It can be understood as a weak
solution in an appropriate sense, but we will not discuss the details here.
The fact that the solution has length scale proportional to t
1/3
after time t
could have been predicted in advance by dimensional analysis, since L = (kIt)
1/3
is the only horizontal length scale in the problem. The numerical factors, and the
fact that the solution has compact support, depend upon the detailed analytical
properties of the porous medium equation equation; they could not be shewn by
dimensional analysis.
8.2. A pedestrian derivation
Let us consider an alternative method for nding the point source solution that
does not require dimensional analysis, but is less systematic.
First, we remove the constants in (2.30) and (2.31) by rescaling the variables.
Dening
u(x,

t) =
1
I
h(x, t) ,

t = kIt,
and dropping the bars on

t, we nd that u(x, t) satises
(2.36) u
t
= (uu
x
)
x
.
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 31
The initial condition (2.31) becomes
u(x, t) 0 as t 0
+
if x ,= 0,
lim
t0
+
_

u(x, t) dx = 1.
(2.37)
We seek a similarity solution of (2.36)(2.37) of the form
(2.38) u(x, t) =
1
t
m
f
_
x
t
n
_
for some exponents m, n. In order to obtain such a solution, the PDE for u(x, t)
must reduce to an ODE for f(). As we will see, this is the case provided that m,
n are chosen appropriately.
Remark 2.12. Equation (2.38) is a typical form of a self-similar solution that is
invariant under scaling transformations, whether or not they are derived from a
change in units. Dimensional analysis of this problem allowed us to deduce that
the solution is self-similar. Here, we simply seek a self-similar solution of the form
(2.38) and hope that it works.
Dening the similarity variable
=
x
t
n
and using a prime to denote the derivative with respect to , we nd that
u
t
=
1
t
m+1
(mf +nf

)
(uu
x
)
x
=
1
t
2m+2n
(ff

.
(2.39)
In order for (2.36) to be consistent with (2.38), the powers of t in (2.39) must
agree, which implies that
(2.40) m+ 2n = 1.
In that case, f() satises the ODE
(ff

+nf

+mf = 0.
Thus, equation (2.36) has a one-parameter family of self-similar solutions. The
ODE for similarity solutions is easy to integrate when m = n, but it is not as
simple to solve when n ,= m.
To determine the value of m, n for the point source problem, we compute that,
for solutions of the form (2.38),
_

u(x, t) dx = t
nm
_

f () d.
Thus, to get a nonzero, nite limit as t 0
+
in (2.37), we must take m = n, and
then (2.40) implies that m = n = 1/3. We therefore recover the same solution as
before.
32
8.3. Scaling invariance
Let us consider the scaling invariances of the porous medium equation (2.36) in
more detail.
We consider a rescaling of the independent and dependent variables given by
(2.41) x = x,

t = t, u = u
where , , are positive constants. Writing u in terms of u in (2.36) and using
the transformation of derivatives

x
=
x
,
t
=

t
,
we nd that u
_
x,

t
_
satises the PDE
u

t
=

2

( u u
x
)
x
.
Thus, the rescaling (2.41) leaves (2.36) invariant if
2
= .
To reformulate this invariance in a more geometric way, let E = R
2
R be the
space with coordinates (x, t, u). For , > 0 dene the transformation
(2.42) g (, ) : E E, g(, ) : (x, t, u)
_
x, t,

2

u
_
.
Then
(2.43) G = g(, ) : , > 0
forms a two-dimensional Lie group of transformations of E:
g(1, 1) = I, g
1
(, ) = g
_
1

,
1

_
,
g (
1
,
1
) g (
2
,
2
) = g (
1

2
,
1

2
)
where I denotes the identity transformation.
The group G is commutative (in general, Lie groups and symmetry groups are
not commutative) and is generated by the transformations
(2.44) (x, t, u)
_
x, t,
2
u
_
, (x, t, u)
_
x, t,
1

u
_
.
Abusing notation, we use the same symbol to denote the coordinate u and the
function u(x, t). Then the action of g(, ) in (2.42) on u(x, t) is given by
(2.45) u(x, t)

2

u
_
x

,
t

_
.
This map transforms solutions of (2.36) into solutions. Thus, the group G is a sym-
metry group of (2.36), which consist of the symmetries that arise from dimensional
analysis and the invariance of (2.36) under rescalings of its units.
8.4. Similarity solutions
In general, a solution of an equation is mapped to a dierent solution by elements
of a symmetry group. A similarity solution is a solution that is mapped to itself
by a nontrivial subgroup of symmetries. In other words, it is a xed point of
the subgroup. Let us consider the case of similarity solutions of one-parameter
subgroups of scaling transformations for the porous medium equation; we will show
that these are the self-similar solutions considered above.
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 33
The one-parameter subgroups of G in (2.43) are given by
H
n
= g (
n
, ) : > 0 for < n < ,
and g (, 1) : > 0. From (2.45), a function u(x, t) is invariant under H
n
if
u(x, t) =
2n1
u
_
x

n
,
t

_
for every > 0. Choosing = t, we conclude that u(x, t) has the form (2.38) with
m = 1 2n. Thus, we recover the self-similar solutions considered previously.
8.5. Translational invariance
A transformation of the space E of dependent and independent variables into it-
self is called a point transformation. The group G in (2.43) does not include all
the point transformations that leave (2.36) invariant. In addition to the scaling
transformations (2.44), the space-time translations
(2.46) (x, t, u) (x , t, u) , (x, t, u) (x, t , u) ,
where < , < , also leave (2.36) invariant, because the terms in the equation
do not depend explicitly on (x, t).
As we will show in Section 9.8, the transformations (2.44) and (2.46) generate
the full group of point symmetries of (2.36). Thus, the porous medium equation
does not have any point symmetries beyond the obvious scaling and translational
invariances. This is not always the case, however. Many equations have point
symmetries that would be dicult to nd without using the theory of Lie algebras.
Remark 2.13. The one-dimensional subgroups of the two-dimensional group of
space-time translations are given by
(x, t, u) (x c, t , u) ,
where c is a xed constant (and also the space translations (x, t, u) (x, t, u)).
The similarity solutions that are invariant under this subgroup are the traveling
wave solutions
u(x, t) = f(x ct).
9. Continuous symmetries of dierential equations
Dimensional analysis leads to scaling invariances of a dierential equation. As we
have seen in the case of the porous medium equation, these invariances form a
continuous group, or Lie group, of symmetries of the dierential equation.
The theory of Lie groups and Lie algebras provides a systematic method to
compute all continuous point symmetries of a given dierential equation; in fact,
this is why Lie rst introduced the the theory of Lie groups and Lie algebras.
Lie groups and algebras arise in many other contexts. In particular, as a result
of the advent of quantum mechanics in the early 20
th
-century, where symmetry
considerations are crucial, Lie groups and Lie algebras have become a central part
of mathematical physics.
We will begin by describing some basic ideas about Lie groups of transforma-
tions and their associated Lie algebras. Then we will describe their application to
the computation of symmetry groups of dierential equations. See Olver [40, 41],
whose presentation we follow, for a full account.
34
9.1. Lie groups and Lie algebras
A manifold of dimension d is a space that is locally dieomorphic to R
d
, although
its global topology may be dierent (think of a sphere, for example). This means
that the elements of the manifold may, locally, be smoothly parametrized by d
coordinates, say
_

1
,
2
, . . . ,
d
_
R
d
. A Lie group is a space that is both a manifold
and a group, such that the group operations (composition and inversion) are smooth
functions.
Lie groups almost always arise in applications as transformation groups acting
on some space. Here, we are interested in Lie groups of symmetries of a dierential
equation that act as point transformations on the space whose coordinates are the
independent and dependent variables of the dierential equation.
The key idea we want to explain rst is this: the Lie algebra of a Lie group
of transformations is represented by the vector elds whose ows are the elements
of the Lie Group. As a result, elements of the Lie algebra are often referred to as
innitesimal generators of elements of the Lie group.
Consider a Lie group G acting on a vector space E. In other words, each g G
is a map g : E E. Often, one considers Lie groups of linear maps, which are a
subgroup of the general linear group GL(E), but we do not assume linearity here.
Suppose that E = R
n
, and write the coordinates of x E as
_
x
1
, x
2
, . . . , x
n
_
.
We denote the unit vectors in the coordinate directions by

x
1,
x
2, . . . ,
x
n.
That is, we identify vectors with their directional derivatives.
Consider a vector eld
v(x) =
i
(x)
x
i ,
where we use the summation convention in which we sum over repeated upper and
lower indices. The associated ow is a one-parameter group of transformations
obtained by solving the system of ODEs
dx
i
d
=
i
_
x
1
, x
2
, . . . , x
n
_
for 1 i n.
Explicitly, if x() is a solution of this ODE, then the ow g() : x(0) x() maps
the initial data at = 0 to the solution at time .
We denote the ow g() generated by the vector eld v by
g() = e
v
.
Conversely, given a ow g(), we can recover the vector eld that generates it from
v(x) =
d
d
g() x

=0
.
That is, v(x) is the tangent, or velocity, vector of the solution curve through x.
Example 2.14. A linear vector eld has the form
v(x) = a
i
j
x
j

x
i .
Its ow is given by the usual exponential e
v
= e
A
where the linear transformation
A has matrix
_
a
i
j
_
.
The ow e
v
of a smooth linear vector eld v is dened for all < < .
The ow of a nonlinear vector eld may exists only for suciently small values of
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 35
, which may depend on the initial data. In that case we get a local Lie group of
ows. Since we only use local considerations here, we will ignore this complication.
9.2. The Lie bracket
In general, a Lie algebra g is a vector space with a bilinear, skew-symmetric bracket
operation
[, ] : g g g
that satises the Jacobi identity
[u, [v, w]] + [v, [w, u]] + [w, [u, v]] = 0.
The Lie bracket of vector elds v, w is dened by their commutator
[v, w] = v w wv,
where the vector elds are understood as dierential operators. Explicitly, if
v =
i

x
i , w =
i

x
i ,
then
[v, w] =
_

j

i
x
j

j

i
x
j
_

x
i .
The Lie bracket of vector elds measures the non-commutativity of the correspond-
ing ows:
[v, w] (x) =
1
2
d
2
d
2
_
e
v
e
w
e
v
e
w
_
x

=0
.
One can show that the Lie bracket of any two vector eld that generate elements
of a Lie group of transformations also generates an element of the Lie group. Thus,
the innitesimal generators of the Lie group form a Lie algebra.
9.3. Transformations of the plane
As simple, but useful, examples of Lie transformation groups and their associated
Lie algebras, let us consider some transformations of the plane.
The rotations of the plane g() : R
2
R
2
are given by
g() :
_
x
y
_

_
cos sin
sin cos
__
x
y
_
where T.
These transformations form a representation of the one-dimensional Lie group
SO(2) on R
2
. They are the ow of the ODE
d
d
_
x
y
_
=
_
y
x
_
.
The vector eld on the right hand side of this equation may be written as
v(x, y) = y
x
+x
y
,
and thus
g() = e
v
.
The Lie algebra so(2) of SO(2) consists of the vector elds
y
x
+x
y
: R .
The translations of the plane in the direction (a, b)
(x, y) (x a, y b)
36
are generated by the constant vector eld
a
x
+b
y
The rotations and translations together form the orientation-preserving Euclidean
group of the plane, denoted by E
+
(2). The full Euclidean group E(2) is generated
by rotations, translations, and reections.
The Euclidean group is not commutative since translations and rotations do
not commute. As a result, the corresponding Lie algebra e(2) is not trivial. For
example, if v =
x
is an innitesimal generator of translations in the x-direction,
and w = y
x
+ x
y
is an innitesimal generator of rotations, then [v, w] =
y
is
the innitesimal generator of translations in the y-direction.
The scaling transformations
(x, y) (e
r
x, e
s
y)
are generated by the vector eld
rx
x
+sy
y
.
Together with the translations and rotations, the scaling transformations generate
the conformal group of angle preserving transformations of the plane.
Finally, as a nonlinear example, consider the vector eld
v(x, y) = x
2

x
y
2

y
.
This generates the local ow
(x, y)
_
x
1 x
,
y
1 +y
_
.
9.4. Transformations of function
Next, we want to consider the action of point transformations on functions.
Suppose that f : R
n
R. We denote the coordinates of the independent vari-
ables by x =
_
x
1
, x
2
, . . . , x
n
_
R
n
, and the coordinate of the dependent variable by
u R. We assume that f is scalar-valued only to simplify the notation; it is straight-
forward to generalize the discussion to vector-valued functions f : R
n
R
m
.
Let E = R
n
R be the space with coordinates
_
x
1
, . . . , x
n
, u
_
. Then the graph

f
of f is the subset of E given by

f
= (x, u) E : u = f(x) .
Consider a local one-parameter group of point transformations g() : E E on
the space of independent and dependent variables. These transformations induce a
local transformation of functions, which we denote in the same way,
g() : f g() f
that maps the graph of f to the graph of g() f. The global image of the graph
of f under g() need not be a graph; it is, however, locally a graph (that is, in a
suciently small neighborhood of a point x R
n
and for small enough values of ,
when g() is suciently close to the identity).
To express the relationship between f and

f = g f explicitly, we write g as
g() : (x, u) ( x, u) , x =

X(x, u, ), u =

U(x, u, ).
Then, since
g() : (x, u) : u = f(x) ( x, u) : u =

f( x, ),
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 37
we have

U (x, f(x), ) =

f
_

X(x, f(x), ),
_
.
This is, in general, a complicated implicit equation for

f in terms of f.
Example 2.15. Suppose that f : R R is the square function f : x x
2
and
g() : R R R R is the rotation
g() (x, u) = ((cos )x (sin )u, (sin )x + (cos )u) .
If u = f(x), then
x = (cos )x (sin )x
2
, u = (sin )x + (cos )x
2
.
Solving the rst equation for x in terms of x, then using the second equation to
express u in terms of x, we nd that u =

f ( x, ) where

f ( x, ) =
2 x
2
cos + 2 xtan
1 2 xsin +
_
1 4 xsin / cos
2

Thus, the image of the function x x

2
under the rotation g() is the function
x
2x
2
cos + 2xtan
1 2xsin +
_
1 4xsin / cos
2

.
Note that this reduces to x x
2
if = 0, and that the transformed function is
only dened locally if ,= 0.
9.5. Prolongation of transformations
In order to obtain the symmetries of a dierential equation, we use a geometric
formulation of how the derivatives of a function transform under point transforma-
tions.
To do this, we introduce a space E
(k)
, called the k
th
jet space, whose coordinates
are the independent variables, the dependent variable, and the derivatives of the
dependent variable of order less than or equal to k.
We will use multi-index notation for partial derivatives. A multi-index is an
n-tuple
= (
1
,
2
, . . . ,
n
)
where each
i
= 0, 1, 2, . . . is a nonnegative integer. The -partial derivative of a
function f : R
n
R is

f =
1
x
1

2
x
2
. . .
n
x
n f.
This partial derivative has order [[ where
[[ =
1
+
2
+ +
n
.
We dene E
(k)
to be the space with coordinates (x, u,

u) where runs over all

multi-indices with 1 [[ k. When convenient, we will use alternative notations
for the partial-derivative coordinates, such as u
x
i for
x
i u and u

for

u.
Example 2.16. Written out explicitly, the coordinates on the rst-order jet space
E
(1)
are
_
x
1
, x
2
, . . . , x
n
, u, u
x
1, u
x
2, . . . , u
x
n
_
. Thus, E
(1)
has dimension (2n + 1).
Example 2.17. For functions u = f(x, y) of two independent variables, the second-
order jet space E
(2)
has coordinates (x, y, u, u
x
, u
y
, u
xx
, u
xy
, u
yy
).
38
Point transformations induce a map of functions to functions, and therefore
they induce maps of the derivatives of functions and of the jet spaces.
Specically, suppose that g() : E E is a point transformation. We extend,
or prolong g(), to a transformation
pr
(k)
g() : E
(k)
E
(k)
in the following way. Given a point (x, u,

u) E
(k)
, pick a function f : R
n
R
whose value at x is u and whose derivatives at x are

u, meaning that
f(x) = u,

f(x) =

u for 1 [[ k.
For example, we could choose f to be a polynomial of degree k.
Suppose that g() (x, u) = ( x, u) is the image of (x, u) E under g() and

f = g()f is the image of the function f. We dene the image of the jet-coordinates
by

u =

f ( x) .
That is, they are the values of the derivatives of the transformed function

f ( x).
One can show that these values do not depend on a particular choice of the function
f, so this gives a well-dened map pr
(k)
g() on E
(k)
such that
pr
(k)
g() : (x, u,

u)
_
x, u,

u
_
.
9.6. Prolongation of vector elds
Suppose that g() : E E is generated by the vector eld
(2.47) v(x, u) =
i
(x, u)
x
i +(x, u)
u
.
Then, writing the coordinates of E
(k)
as (x, u, u

), the prolonged transformation

pr
(k)
g() : E
(k)
E
(k)
is generated by a vector eld pr
(k)
v on E
(k)
. This prolonged vector eld has the
form
pr
(k)
v =
i

x
i +
u
+
k

||=1

u
,
where the

are suitable coecient functions, which are determined by v.

The prolongation formula expresses the coecients

of the prolonged vector

eld in terms of the coecients
i
, of the original vector eld. We will state the
result here without proof (see [40] for a derivation).
To write the prolongation formula in a compact form see (2.49) below
we dene the total derivative D
x
i F : E
(k)
R of a function F : E
(k)
R with
respect to an independent variable x
i
by
D
x
i F =
x
i F +
k

||=0
u
,i

u
F.
Here, we use the notation
u
,i
=
x
i

u
to denote the coordinate of the corresponding derivative. That is, u
,i
= u

where

i
=
i
+ 1 and
j
=
j
for j ,= i.
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 39
In other words, the total derivative D
x
i F of F with respect to x
i
is what we
would obtain by dierentiating F with respect to x
i
after the coordinates u, u

have been evaluated at a function of x and its derivatives.

If = (
1
, . . . ,
n
) is a multi-index, we dene the -total derivative by
D

= D
1
x
1
D
2
x
2
. . . D
n
x
n .
Total derivatives commute, so the order in which we take them does not matter.
Finally, we dene the characteristic Q : E
(1)
R of the vector eld (2.47) by
Q(x, u, u) = (x, u)
i
(x, u)u
x
i ,
where the summation convention is understood, and u = (u
x
1, . . . , u
x
n) is the
rst-derivative coordinate.
Then the k
th
-prolongation of the vector eld (2.47) is given by
pr
(k)
v =
i

x
i +
u
+
k

||=1

u
, (2.48)

= D

Q+
i
u
,i
. (2.49)
This is the main result needed for the algebraic computation of symmetries of a
dierential equation. See Olver [40] for the prolongation formula for systems.
9.7. Invariance of a dierential equation
A k
th
order dierential equation for a real-valued function u(x) may be written as
F (x, u,

u) = 0
where F : E
(k)
R and 1 [[ k. Here, we abuse notation and use the same
symbols for the coordinates u,

u and the functions u(x),

u(x).
A local point transformation g() : E E is a symmetry of the dierential
equation if g() u is a solution whenever u is a solution. This means that, for all
in the neighborhood of 0 for which g() is dened, we have
(2.50) F
_
pr
(k)
g() (x, u,

u)
_
= F (x, u,

u) on F (x, u,

u) = 0.
Suppose that g() = e
v
. Then, dierentiating (2.50) with respect to and
setting = 0, we conclude that
(2.51) pr
(k)
v F (x, u,

u) = 0 on F (x, u,

u) = 0,
where pr
(k)
v acts on F by dierentiation. Conversely, if F satises the maximal
rank condition DF ,= 0 on F = 0, which rules out degenerate ways of the equation
such as F
2
= 0, we may integrate the innitesimal invariance condition (2.51) to
obtain (2.50).
The condition (2.51) is called the determining equation for the innitesimal
symmetries of the dierential equation. It is typically a large, over-determined sys-
tem of equations for
i
(x, u), (x, u) and their derivatives, which is straightforward
(though tedious) to solve.
Thus, in summary, to compute the point symmetries of a dierential equation
F (x, u,

u) = 0
we use the prolongation formula (2.48)(2.49) to write down the innitesimal in-
variance condition (2.51), solve the resulting equations for
i
(x, u) and (x, u), then
integrate the vector elds (2.47) to obtain the symmetries g = e
v
.
40
9.8. Porous medium equation
Let us return to the porous medium equation (2.36).
The space E of independent and dependent variables has coordinates (x, t, u).
We may write the equation as
(2.52) F (u, u
x
, u
t
, u
xx
) = 0
where F : E
(2)
R is given by
F (u, u
x
, u
t
, u
xx
) = u
t
+u
2
x
+uu
xx
.
A vector eld v on E is given by
(2.53) v(x, t, u) = (x, t, u)
x
+(x, t, u)
t
+(x, t, u)
u
.
From (2.48), the second prolongation of v has the form
pr
(2)
v =
x
+
t
+
u
+
x

ux
+
t

ut
+
xx

uxx
+
xt

uxt
+
tt

utt
.
The innitesimal invariance condition (2.51) applied to (2.52) gives
(2.54)
t
+ 2u
x

x
+u
xx
+u
xx
= 0 on u
t
= u
2
x
+uu
xx
.
From (2.49), we have

x
= D
x
Q+u
xx
+u
xt
,

t
= D
t
Q+u
xt
+u
tt
,

xx
= D
2
x
Q+u
xxx
+u
xxt
,

xt
= D
x
D
t
Q+u
xxt
+u
xtt
,

tt
= D
2
t
Q+u
xtt
+u
ttt
,
where the characteristic Q of (2.53) is given by
Q(x, t, u, u
t
, u
x
) = (x, t, u) (x, t, u) u
x
(x, t, u) u
t
,
and the total derivatives D
x
, D
t
of a function f (x, t, u, u
x
, u
t
) are given by
D
x
f = f
x
+u
x
f
u
+u
xx
f
ux
+u
xt
f
ut
,
D
t
f = f
t
+u
t
f
u
+u
xt
f
ux
+u
tt
f
ut
.
Expanding the total derivatives of Q, we nd that

x
=
x
+ (
u

x
) u
x

x
u
t

u
u
2
x

u
u
x
u
t
,

t
=
t

t
u
x
+ (
u

t
) u
t

u
u
x
u
t

u
u
2
t
,

xx
=
xx
+ (2
xu

xx
) u
x

xx
u
t
+ (
uu
2
xu
) u
2
x
2
xt
u
x
u
t

uu
u
3
x

uu
u
2
x
u
t
+ (
u
2
x
) u
xx
2
x
u
xt
3
u
u
x
u
xx

u
u
t
u
xx
2
u
u
x
u
xt
.
We use these expressions in (2.54), replace u
t
by uu
xx
+u
2
x
in the result, and equate
coecients of the terms that involve dierent combinations of the spatial derivatives
of u to zero.
The highest derivative terms are those proportional to u
xt
and u
x
u
xt
. Their
coecients are proportional to
x
and
u
, respectively, so we conclude that
x
= 0
and
u
= 0, which implies that = (t) depends only on t.
The remaining terms that involve second-order spatial derivatives are a term
proportional to u
x
u
xx
, with coecient
u
, so
u
= 0 and = (x, t), and a term
LECTURE 2. DIMENSIONAL ANALYSIS, SCALING, AND SIMILARITY 41
proportional to u
xx
. Equating the coecient of the latter term to zero, we nd
that
(2.55) = (2
x

t
) u.
Thus, is a linear function of u.
The terms that are left are either proportional to u
2
x
, u
x
, or involve no deriva-
tives of u. Equating to zero the coecients of these terms to zero, we get

t
2
x
+
u
+u
uu
= 0,

t
u
xx
+ 2
x
+ 2u
xu
= 0,

t
u
xx
= 0.
The rst equation is satised by any of the form (2.55). The second equation
is satised if
t
= 0 and
xx
= 0, which implies that
=
1
+
3
x
for arbitrary constants
1
,
3
. The third equation holds if
tt
= 0, which implies
that
=
2
+
4
t
for arbitrary constants
2
,
3
. Equation (2.55) then gives
= (2
3

4
) u
Thus, the general form of an innitesimal generator of a point symmetry of
(2.36) is
v(x, t, u) = (
1
+
3
x)
x
+ (
2
+
4
t)
t
+ (2
3

4
) u
u
.
We may write this as
v =
4

i=1

i
v
i
where the vector elds v
i
are given by
v
1
=
x
, v
2
=
t
(2.56)
v
3
= x
x
+ 2u
u
v
4
= t
t
u
u
(2.57)
The vector elds v
1
, v
2
in (2.56) generate the space and time translations
(x, t, u) (x +
1
, t, u), (x, t, u) (x, t +
2
, u),
respectively. The vector elds v
3
, v
4
in (2.57) generate the scaling transformations
(x, t, u)
_
e
3
x, t, e
23
u
_
(x, t, u)
_
x, e
4
t, e
4
u
_
.
These are the same as (2.44) with
= e
3
, = e
4
.
Thus the full point symmetry group of the porous medium equation is four
dimensional, and is generated by space and time translations and the two scaling
transformations that arise from dimensional analysis.
This result is, perhaps, a little disappointing, since we did not nd any new
symmetries, although it is comforting to know that there are no other point sym-
metries to be found. For other equations, however, we can get symmetries that are
not at all obvious.
42
Example 2.18. Consider the one-dimensional heat equation
(2.58) u
t
= u
xx
.
The determining equation for innitesimal symmetries is

t
=
xx
on u
t
= u
xx
.
Solving this equation and integrating the resulting vector elds, we nd that the
point symmetry group of (2.58) is generated by the following transformations [40]:
u(x, t) u(x , t),
u(x, t) u(x, t ),
u(x, t) u(x, t),
u(x, t) u(x,
2
t),
u(x, t) e
x+
2
t
u(x 2t, t),
u(x, t)
1

1 + 4t
exp
_
x
2
1 + 4t
_
u
_
x
1 + 4t
,
t
1 + 4t
_
,
u(x, t) u(x, t) +v(x, t),
where (, . . . , ) are constants, and v(x, t) is an arbitrary solution of the heat equa-
tion. The scaling symmetries involving and can be deduced by dimensional
arguments, but the symmetries involving and cannot.
As these examples illustrate, given a dierential equation it is, in principle,
straightforward (but lengthy) to write out the conditions that a vector eld gen-
erates a point symmetry of the equation, solve these conditions, and integrate the
resulting innitesimal generators to obtain the Lie group of continuous point sym-
metries of the equation. There are a number of symbolic algebra packages that will
do this automatically.
Finally, we note that point symmetries are not the only kind of symmetry
one can consider. It is possible to dene generalized (also called nonclassical
or higher) symmetries on innite-dimensional jet spaces (see [40], for an intro-
duction). These are of particular interest in connection with completely integrable
equations, such as the Korteweg-de Vries (KdV) equation, which possess hidden
symmetries that are not revealed by their point symmetries